Summary of Hands-on Impressions: Using Codex and Open Code for Personal Use while using Claude Code for Work
Hello, I'm Munou.
I feel like there aren't many comprehensive summaries out there, so I'm going to write down my thoughts (as fairly as possible) after using various tools. There's a lot of position talk.
First of all, is it even possible to evaluate them fairly when their use cases are different?
Impressions of Each Model
Codex
My main tool is usually Codex.
The intelligence in suggesting bug fixes and implementing them is characteristic of OpenAI's LLM models, and because it runs on a powerful infrastructure, the response speed is fast and not frustrating. It handles things with minimal fixes, making it easy to read with git diff, though other LLM models are getting smarter too, so that might not be as true anymore.
Usually, if I provide a problematic debug log, it can mostly solve it, and I rarely get stuck.
Personally, the review capability of the web version of ChatGPT is powerful, and in my personal environment, I often go back and forth between ChatGPT reviews and Codex. Since I can reflect the results of brainstorming directly into the code, it reduces wasted commits.
Plus, it doesn't consume the Codex quota, so I recommend using it alongside the web version of ChatGPT.
My personal guess is that they might increase reset opportunities to pump up the valuation before the IPO, but currently, token consumption from going back and forth with the web version isn't decreasing at all, so it's not a problem.
I assume multi-agent usage isn't natively supported, so in that case, it might be better to use other coding agents.
Open Code
I use it for simple tasks or research, like making throwaway bash scripts, investigating raw logs directly, or applying simple GitHub Actions to multiple repositories. I mainly use Xiaomi's Mimo V2.5 model, which is still free, and its efficiency when running research tasks with multi-agents is quite good.
I also have it summarize actual API endpoint response investigations and then implement them on the Codex side. Since it learns, I revoke API keys immediately in those cases.
My security awareness is loose, but it's simpler and better than doing roundabout things to protect API keys.
I don't really use it as my main tool for coding.
Regarding the quota, it seems to be a 9H quota, but I've only hit the limit once, so I think the quota is quite generous for a free tier.
However, for coding, I'm currently using GLM-5.2 via Neuralwatt, and it's useful because it finds security-related issues that even Codex misses due to its overwhelming context length. I feel GLM-5.2 is quite practical for coding beyond just security tasks. I haven't used Kimi, so I don't know.
On the negative side, if I repeat opening and closing it many times, it puts a lot of temporary files in /tmp/, and before I know it, it won't start.
Claude Code
It outputs a lot overall, so it floods. It might be good for initial app coding, but I'm not sure if I want to use it for personal development. It sometimes implements more than necessary, which can be troublesome.
Security investigation tasks are currently useless because Fable's censorship is too severe.
And I think this was the same for the personal plan, but it's treated the same as the web chat quota, so if you hit the limit, you're done. However, I think business quotas for ChatGPT/Codex were also separate, so it might be effectively the same.
However, my favorite developer, xroche, has resumed maintenance of HTTrack (which hadn't been maintained for a while) using Claude, so I think my negative impression is just because I've gotten used to Codex.
Another negative image is that while Anthropic is good at marketing as a business, I sometimes feel there's a gap with reality. It might be something they're doing to pump up valuation before the IPO, but they engage in activities that make me wonder, such as:
Activities for policies to slow down China's AI development speed
Activities to prevent unauthorized use of computing resources by Chinese companies
These might function as a spokesperson for activities to protect US industry, but I feel some sympathy for the regulations related to Chinese semiconductors, and I question the sustainability or political clumsiness.
Also, the fact that FTX / Alameda held shares in the past is a bit unsettling. It's already been settled through bankruptcy and was just one of many investments, but it leaves me with an indescribable feeling.
Conclusion
Honestly, most models are excellent now, so in terms of implementation, I don't think there's much difference if used properly, but I really want people who are just getting by with Open Code to try Codex once.
That's all.